GeoWeaver couples a geometric prior model with test-time adaptation to fix long-video 3D reconstruction drift

zhenjun_zhao · x · 2026-08-19

New arXiv paper (2608.17389, Tinghao Jiang et al.). Long-sequence RGB video 3D reconstruction needs both accurate local geometry and globally consistent camera motion; feed-forward models predict strong depth and pose but are memory-bound, while chunk-wise processing causes inter-chunk scale drift, pose errors, and point-cloud misalignment.

GeoWeaver comprises a Geometric Prior Model (GPM) that predicts chunk-wise depth, confidence, and camera parameters as adjustable priors, and Test-Time Adaptation (TTA) that performs sequential initialization, global chunk-level Sim(3) alignment, and coarse-to-fine refinement of camera poses, affine depth corrections, and intrinsics. Dense correspondences provide adjacent, cross-chunk, and long-range constraints, while a robust CDF-style objective jointly optimizes weighted 2D reprojection and 3D consistency residuals. Experiments across long-sequence benchmarks show improved camera accuracy, global consistency, and point-cloud quality; ablations verify each adaptation stage, and the same TTA procedure transfers to other feed-forward models.

Original post →

More from Multimodal

Multimodal channel →